How to Keep Confidential Information Out of AI Tools

When confidential information ends up somewhere it should not, it is usually not a break-in. It is a busy person pasting a whole email thread into a chat box to get a quick summary. The fix is not to scare people. It is to make the safe choice the easy one.

Start by naming what "confidential" means here

Telling staff "do not put sensitive data into AI" does not work, because everyone draws the line in a different place. Write it down in plain words, with examples from your own work. A short list beats a long definition.

• People. Names, contact details, birth dates, ID numbers, health information, case notes, and anything about a client, patient, student, or employee.
• Access. Passwords, access keys, account numbers, and security settings.
• Agreements. Anything covered by a confidentiality agreement or a contract that limits how it can be shared.
• Legal and personnel matters. Complaints, investigations, discipline, and anything involving an attorney.
• Unreleased work. Bids, pricing, plans, and drafts that have not been made public.

Then put the list where people will see it when they need it: in your AI use policy, in onboarding, and next to the tools themselves.

Know which tools are approved for what

"Is AI allowed?" is the wrong question. The useful one is "which tool, for which kind of information?" A tool your organization has reviewed and set up under a business agreement may be fine for some internal material. A personal account on a free tool almost never is.

Keep a simple table: tool, who may use it, and what kind of information may go into it. If a tool is not on the table, it is not approved for anything confidential. If you are not sure what is already in use, read what to do when staff are already using AI without approval first.

Teach people to strip it before they paste

Most work requests do not need the confidential details at all. If someone wants help rewording a letter, the tool does not need the client's name, address, or account number to do that.

1. Replace names with roles. "The client," "the employee," "the vendor."
2. Take out numbers that identify. Account numbers, case numbers, ID numbers, addresses, and phone numbers.
3. Cut what the task does not need. If you are asking for a better opening paragraph, paste the opening paragraph, not the whole thread.
4. Read what is left. Could someone who knows the situation still tell who it is about? A title, a date, and an unusual detail together can give a person away.
5. Put the details back yourself in the final version, outside the tool.

This takes a minute. It is a habit, and habits come from practice, not from a policy someone signed once.

Watch the places it slips in

People are careful with the obvious things. The leaks tend to come from the less obvious ones:

• Forwarded email threads, where the confidential part is six replies down.
• Uploaded files, where a spreadsheet has a hidden tab or a document has comments and tracked changes.
• Screenshots, which carry whatever else was on the screen.
• Meeting recordings and transcripts sent to a summarizing tool, with everything anyone said in the room.
• Browser extensions and add-ons that read the page you are on. Ask what they can see before you install them.
• AI features turned on inside software you already use. A new button in your email or document program may send content somewhere new. Find out before you click it.

Set the tools up so the default is safe

Training helps, but settings do more of the work. Whoever administers your approved tools should check:

• Whether your content is used to train or improve the vendor's models, and whether that is switched off for every user, not just the administrator.
• How long prompts, files, and outputs are kept, and whether you can shorten that.
• Whether people sign in with their work account, so access ends when they leave.
• Whether there is a log of who used the tool, so a problem can be traced.

Write down what you checked and when. Settings and terms change, so put a date on the calendar to check again.

When it happens anyway

Someone will paste the wrong thing at some point. What matters is whether they tell you. If the first person to admit a mistake gets punished for it, the next person will hide theirs, and you will find out much later, from someone else.

Make the steps simple: stop, write down the tool, the account, the date, and what was shared, and tell the person who handles privacy or security that same day. Some kinds of information come with legal duties when they are exposed, and your organization can only meet them if it knows. That call belongs to your privacy lead or attorney, not to the person who made the mistake. Our guide to what to do when AI gets something wrong at work covers the reporting habit in more detail.

The short version for your team

• Use only the tools on the approved list, for the information they are approved for.
• Strip names, numbers, and details the task does not need.
• Check attachments, threads, and screenshots before they go in.
• If you are not sure, leave it out and ask.
• If something slipped through, say so the same day.

Five lines. Short enough to remember, which is the whole point.